Journal of Clinical Epidemiology 66 (2013) 752e758

Stepped wedge designs could reduce the required sample size

in cluster randomized trials
Willem Woertmana,1, Esther de Hoopa,*,1, Mirjam Moerbeekb, Sytse U. Zuidemac,
Debby L. Gerritsend, Steven Teerenstraa
a
Department for Health Evidence, Radboud University Medical Centre, 133, PO Box 9101, 6500 HB, Nijmegen, The Netherlands
b
Department of Methodology and Statistics, Utrecht University, PO Box 80089, 3508 TB, Utrecht, The Netherlands
c
Department of General Practice, University of Groningen, University Medical Centre Groningen, PO Box 72, 9700 AB, Groningen, The Netherlands
d
Department of Primary and Community Care, Centre for Family Medicine, Geriatric Care and Public Health, Radboud University Medical Centre, 117,
PO Box 9101, 6500 HB, Nijmegen, The Netherlands
Accepted 8 January 2013; Published online 22 March 2013

Abstract
Objective: The stepped wedge design is increasingly being used in cluster randomized trials (CRTs). However, there is not much
information available about the design and analysis strategies for these kinds of trials. Approaches to sample size and power calculations
have been provided, but a simple sample size formula is lacking. Therefore, our aim is to provide a sample size formula for cluster
randomized stepped wedge designs.
Study Design and Setting: We derived a design effect (sample size correction factor) that can be used to estimate the required sample
size for stepped wedge designs. Furthermore, we compared the required sample size for the stepped wedge design with a parallel group and
analysis of covariance (ANCOVA) design.
Results: Our formula corrects for clustering as well as for the design. Apart from the cluster size and intracluster correlation, the design
effect depends on choices of the number of steps, the number of baseline measurements, and the number of measurements between steps.
The stepped wedge design requires a substantial smaller sample size than a parallel group and ANCOVA design.
Conclusion: For CRTs, the stepped wedge design is far more efficient than the parallel group and ANCOVA design in terms of sample
size.
Ó 2013 Elsevier Inc. Open access under the Elsevier OA license.

Keywords: Cluster randomized trial; Stepped wedge design; Parallel group design; ANCOVA; Sample size; Design effect

1. Introduction The most commonly used trial design is the parallel

group design in which each cluster is randomized to either
Randomized controlled trials are considered the gold
an intervention or control condition [4]. Within this design,
standard in evaluating health care interventions [1]. How-
each cluster receives only one kind of treatment during
ever, cluster randomized trials (CRTs) are increasingly
the study, and usually all clusters start simultaneously.
being used in the health care setting [2]. In these trials, An extension of the parallel group design is the analysis
complete social units, or groups of individuals (such as
of covariance (ANCOVA) design where a baseline mea-
families, nursing homes [NHs], or general practices), are
surement is added to the design and included as a covariate
randomized to different treatments. They are mostly used
in the analysis [5]. In contrast, in the crossover design, ev-
to prevent contamination and in situations where individual
ery cluster will receive both the intervention and the control
randomization is not possible or not desirable for logistic,
treatment. Yet, the order of the interventions is randomized
financial, or ethical reasons [3].
for each cluster [3,4,6]. However, it is not always possible
to conduct a crossover design because it assumes that the
carryover effects are absent [3,4,6]. This means that the
Conflict of interest: Neither the article nor any parts of it have been estimated treatment effects should be independent of the
published or submitted before. No external funding has been received order in which the treatments were assigned. So, the effects
and no conflict of interest is present.
1
W.W and E.de.H. are the joint contributions.
of the first treatment should have disappeared by the time
* Corresponding author. Tel.: þ31-24-3667262; fax: þ31-24-3613505. the second treatment is started, which may be unrealistic
E-mail address: E.deHoop@ebh.umcn.nl (E. de Hoop). if, for example, the first treatment is the reinforcement of
0895-4356 Ó 2013 Elsevier Inc. Open access under the Elsevier OA license.
http://dx.doi.org/10.1016/j.jclinepi.2013.01.009
 W. Woertman et al. / Journal of Clinical Epidemiology 66 (2013) 752e758
What is new?
 Approaches to power calculations for cluster ran-
domized stepped wedge designs have been pro-
vided, but a simple sample size formula is
lacking. Therefore, we present a sample size for-
mula for these kinds of trials.
 We derived a formula in which, besides the cluster
size and intracluster correlation, the number of
steps and measurements can be varied.
 The stepped wedge design requires a substantial
smaller sample size than a parallel group or analy-
sis of covariance design.
a hygiene protocol and the second implies falling back to
usual care in hospital wards.
Herein, we will focus on the stepped wedge design,
which is a type of crossover design in which (different)
clusters switch treatments in only one direction at different
time points (steps) [7e10]. Typically, all clusters start in the
control condition. Then, the clusters switch to the interven-
tion at consecutive time points, where the time of the
switch is randomized for every cluster. Eventually, all clus-
ters will have switched from one condition to the other
(see Fig. 1).
The stepped wedge design is especially useful when the
intervention is thought to do more good than harm (i.e.,
when there is no equipoise) [8e10]. In that situation, it is
unethical to withhold or withdraw the intervention from
a proportion of the subjects as would occur in a parallel
group or crossover design, respectively. Besides, it may
be impossible to implement the intervention in half of all
clusters simultaneously because of practical, logistical, or
financial reasons [8e10]. Then, the stepwise treatment im-
plementation of the stepped wedge design offers a solution.
Control
3
Treatment
2 design was chosen. First, the number of available units was
1
Fig. 1. Illustration of the stepped wedge design, where different
(groups of) clusters switch from control to treatment at different
time points.
753
In addition, there are other advantages of the stepped
wedge design. First, the clusters act as their own controls
because they receive both the control and treatment condi-
tions. Therefore, the intervention effect can be estimated
from both between- and within-cluster comparisons. This
results in more statistical power and smaller required sam-
ple sizes than in a parallel group design [8]. Furthermore, it
is possible to control for time with the stepped wedge de-
sign [9]. By modeling the effects of time, it is possible to
study whether the time spent in the intervention condition
influences the effectiveness of the treatment. Finally, re-
cruitment of clusters and/or subjects may be easier within
this design because everyone will receive the treatment dur-
ing the trial.
In this article, we present a relatively simple sample size
formula for stepped wedge CRTs. A recent review showed
that the stepped wedge design is increasingly being used
over the last couple of years [10]. Yet, it was noted that
the reporting of stepped wedge CRTs needs to be improved,
especially the reporting of sample size and power calcula-
tions. Hussey and Hughes [8] provide approaches to sample
size and power calculations. However, their approach does
not provide a sample size formula. Therefore, we propose
a simpler sample size approach using a design effect (sam-
ple size correction factor).
In Section 2, we describe a trial in which the stepped
wedge design is being used. Throughout the article, we will
use this trial as an example. In Section 3, a sample size for-
mula will be presented, and a comparison with the parallel
group and ANCOVA design will be made in Section 4. We
will conclude with a summary and discussion in Section 5.
2. Exampledthe Act in Case of Depression study
Depression is a common health problem in NH resi-
dents. However, it is often undetected and undertreated.
Therefore, the Nijmegen University Network of Nursing
Homes developed the Act in Case of Depression (AiD)
program [11]. This is a multidisciplinary care program to
identify and treat depression, and to monitor the treatment
effects. Because the AiD program involves the training and
cooperation of nursing staff, physicians, psychologists, and
recreational therapists in the NHs, this program is naturally
implemented at the unit level (ward) of the NHs.
The AiD study is a CRT using a stepped wedge design
that aims to assess the efficacy of the AiD program in
NH units. There are two main reasons why a stepped wedge
small. Therefore, a parallel group design would not have
sufficient power to detect a relevant treatment effect (see
Section 4). Second, it was impractical to implement the
program in half of the participating units simultaneously
because of the substantial training effort that was required.
Hence, stepwise implementation of the program was pre-
ferred. Obviously, a crossover design was impossible for
754 W. Woertman et al. / Journal of Clinical Epidemiology 66 (2013) 752e758
this trial because the training of all professionals involved
could not be undone.
3. Sample size calculations
In this section, we will provide a relatively simple sam-
ple size formula for CRTs with a stepped wedge design.
This formula is derived from the formulae provided by
Hussey and Hughes [8]. Their work is based on a model
that comes with the following assumptions that will hold
for our formula as well. First, it is assumed that there are
random cluster effects, fixed time effects, and absence of
cluster by time interactions. That is, the variation of a clus-
ter mean over time is only because of changing subjects
over time, and there is no inherent variation at the cluster
level over time. Furthermore, the model takes external time
trends into account. Yet, these trends are assumed to be
equal for all clusters. Second, it is assumed that there is
no within-subject correlation over time. This is likely to
hold if different subjects are sampled from the clusters at
each measurement point. In contrast to Hussey and Hughes
[8], we consider a simpler situation where the same number
of clusters switches at each step and where the number of
measurements after each step is constant as well.
The standard approach for sample size calculations for
parallel group CRTs is to calculate the sample size that
would be needed if individuals were to be randomized
(Nu). Then, this unadjusted sample size is multiplied by
the design effect ½1 þ ðn  1Þr to correct for clustering,
where n is the number of subjects within a cluster and
r is the intracluster correlation (ICC) [3]. For an ANCOVA
design, the sample size for a clustered parallel group design
nr
is multiplied by a factor ð1  r2 Þ; where r5
1 þ ðn  1Þr
(derived from [5] with cluster autocorrelation set to 1 and
subject autocorrelation set to 0 to obtain the same model
as Hussey and Hughes [8]).
To be able to use a similar approach for stepped wedge
designs, we derived the following design effect:
1 þ rðktn þ bn  1Þ 3ð1  rÞ
DEsw 5  $  
1 1
1 þ r ktn þ bn  1 2t k 
2 k
where k is the number of steps, b is the number of baseline
measurements, and t is the number of measurements after
each step. Hence, the clusters will be measured b þ k$t
times each. This design effect corrects for both clustering
and the stepped wedge design. See Appendix at www.
jclinepi.com for derivations.
Now, the required sample size for the stepped wedge
design can be calculated by multiplying the unadjusted
sample size by the design effect: Nsw 5Nu $DEsw . Note that
formulae for Nu usually result in the number of subjects
per treatment arm, whereas the total number of subjects
is needed here. The required number of clusters c is
calculated by dividing the required sample size Nsw by
the cluster size n, and the number of clusters switching
treatment at each step is calculated by dividing the number
of clusters c by the number of steps k. Obviously, it is not
guaranteed that the required number of clusters will be an
integer. If not, round this number off to the integer above.
The same holds for the number of clusters that should
switch at every step. However, it suffices to distribute the
clusters as evenly as possible over the steps.
The design effect DEsw is affected by choices regarding
three determinants. To guide the choice of these determi-
nants, we will describe how each of them changes the
design effect and hence the sample size. First, it can be seen
that the design effect decreases as the number of measure-
ments t after each step increases. The same holds for the
number of baseline measurements b and the number of
steps k. That is, increasing the number of steps or the num-
ber of baseline measurements decreases the design effect.
So, increasing one or more of the above-mentioned three
determinants decreases the required sample size. However,
increasing the cluster size n results in a slightly larger
design effect.
Besides the previously described design choices, the
design effect also depends on the variation between clus-
ters, that is, the ICC r. The ICC cannot be chosen freely
but depends on the context of the trial (e.g., population of
subjects, type of clusters, and type of outcome measure).
Reasonable estimates for the ICC need to be motivated
by previous comparable studies, for example, pilot studies
or context matter knowledge. As the ICC increases, the de-
sign effect first slightly increases as well (up to about
r50:05Þ, and then starts decreasing. Figure 2 shows the
effect of the determinants on the design effect and hence
the sample size. Furthermore, Table 1 shows the required
number of clusters for a parallel group, ANCOVA, and
stepped wedge design given the several effect sizes and
numbers of steps.
Various choices for the determinants of the design effect
will lead to different study designs, which may differ in the
required sample size. For example, suppose the number of
clusters is fixed at 12. Then, it is possible to opt for a design
with two steps, both at which six clusters will switch. How-
ever, other options are to choose for 3, 4, or 6 steps at which
4, 3, or 2 clusters will switch, respectively. The most
extreme variant would be a design at which only one cluster
switches at each step, which results in 12 steps. Then, the
question arises which design will require the smallest num-
ber of subjects. If t is fixed, a design with 12 steps will lead
to the smallest number of subjects. This is rather straight-
forward because the total number of measurements is
b þ 12$t for this option, which is larger than b þ 2$t mea-
surements for a design with two steps, for example. There-
fore, more information will be available and hence fewer
subjects are required.
In general, it can be shown that efficiency in terms of
sample size improves if the number of steps increases.
 W. Woertman et al. / Journal of Clinical Epidemiology 66 (2013) 752e758 755
1.4

1.4

1.4
1.2

1.2

1.2
1.0

1.0

1.0
0.8

0.8

0.8
0.6

0.6

0.6
0.4

0.4

0.4
0.2

0.2

0.2
0.0

0.0

0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20 0.0 0.00 0.05 0.10 0.15 0.20

Fig. 2. The design effect for a stepped wedge design as a function of the intracluster correlation for various cluster sizes (n), numbers of steps (k),
and measurements after each step (t). The number of baseline measurements (b) equals t. From top to bottom, the long-dashed lines represent
k 5 2, dotted lines represent k 5 3, dashed lines represent k 5 5, and solid lines represent k 5 10.

However, the efficiency gain from going from 2 to 3 steps is the number of clusters is known in advance and the ques-
much larger than the efficiency gain from going from 6 to tion is how many subjects per cluster should be sampled.
0
12 steps. Moreover, the efficiency gain of increasing the In this situation, guess a reasonable cluster size ðn Þ, calcu-
0
number of steps while keeping the number of clusters and late the number of clusters needed (c ), and compare this
the total number of measurements fixed is modest com- number with the available number of clusters ðcÞ. If
0
pared with the efficiency gain from adding clusters or mea- c0 !c, the cluster size n can be decreased, whereas the
surements to the design. cluster size should be increased if c0 Oc. This process
0
Within the presented sample size approach, the number should be repeated until a cluster size n is found for which
0
of subjects per cluster is chosen in advance and the required c is close to or equals c, but not larger than c.
number of clusters is calculated. However, it may be that Although this method is rather straightforward, it should
be noted that, for CRTs in general, it may be difficult or
Table 1. Total number of clusters needed for a parallel group design impossible to obtain sufficient power only by increasing
(PGD), ANCOVA, or stepped wedge design with several numbers cluster sizes, especially if the number of clusters is small
of steps (k) for several effect sizes [12e16].
Effect size 0.2 0.3 0.4 0.5
PGD 114 51 29 19 Example. The AiD program was expected to reduce the
ANCOVA 101 45 26 17 prevalence of depression in psychogeriatric NH units from
Stepped wedge 30% to 19.5%. Uncorrected for clustering and repeated
With k 5 2 94 42 24 15 measurements, a total sample size (Nu ) of 598 residents
With k 5 3 57 25 15 9
With k 5 5 35 16 9 6
would be required to detect this effect (power of 80% and
significance level of 0.05). It was expected that the number
Abbreviation: ANCOVA, analysis of covariance.
Note: Cluster size n 5 10, intracluster correlation 5 0.05, and the
of participating patients per NH unit would be around 20
number of baseline measurements and measurements after each (n). Furthermore, the ICC was estimated to be at most r
step b 5 t 5 1.  0:10 from a pilot study. The total study period was 24
756 W. Woertman et al. / Journal of Clinical Epidemiology 66 (2013) 752e758

months, and it was expected that at least 4 months would be
needed for training and implementation of the AiD pro-
gram. Therefore, at most, six measurements could be taken.
We chose to maximize the number of steps, so k 5 5.
Hence, one baseline measurement (b 5 1) and one mea-
surement after each step (t 5 1) had to be taken. This re-
sulted in a design effect of DEsw 5 0.46 and therefore
a total sample size of 275 patients. So, the total sample size
was reduced by more than 50% in comparison to a study
where individuals would be randomized and measured only
once (after the intervention). Dividing the total sample size
by the cluster size resulted in 14 clusters of which 2.8
should switch at every step. Hence, three clusters would
switch simultaneously at four of the steps and two clusters
at the remaining step.
4. Comparing stepped wedge with parallel group and
ANCOVA design
We presented sample size formulae for the stepped
wedge, ANCOVA, and parallel group design in CRTs. All
use a design effect approach, where the design effect for
a stepped wedge design is presented previously, the design
1.0
effect for an ANCOVA design is ð1  r 2 Þ  ½1 þ ðn  1Þr,
and the design effect for a parallel group design is
½1 þ ðn  1Þr. Therefore, the efficiency of these designs
can be compared by dividing the design effect for a stepped
wedge and ANCOVA design by the one for a parallel group
design. Figure 3 shows this ratio for several design choices.
It can be seen that the stepped wedge design is more effi-
cient than a parallel group design, and its reduction in
sample size becomes more pronounced when the ICC, clus-
ter size, and number of steps and measurements increase.
However, even in a design with one measurement after each
step, a small cluster size and an ICC of 0.05, the stepped
wedge design with three steps already reduces the sample
size with more than 40% in comparison to a parallel group
design. Besides, it can be seen that the ANCOVA design is
also more efficient than the parallel group design, but less
efficient than the stepped wedge design.
Example. Previously, we showed that the AiD trial using
a stepped wedge design with five steps required 275 patients.
For a cluster randomized parallel group design, the design
effect is DEpgd 51 þ ðn  1Þr51 þ ð20  1Þ  0:1052:9.
Multiplying this design effect with the unadjusted sample
size of 598 patients results in a total of 1,735 patients. So,

1.0

1.0
0.8

0.8

0.8
0.6

0.6

0.6
0.4

0.4

0.4
0.2

0.2

0.2
0.0

0.0

0.0

0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20

Fig. 3. Efficiency of an analysis of covariance (solid line) and the stepped wedge design relative to a parallel group cluster randomized trial as
a function of the intracluster correlation for various cluster sizes (n), numbers of steps (k), and measurements after each step (t). The number
of baseline measurements (b) equals t. From top to bottom, the long-dashed lines represent k 5 2, dotted lines represent k 5 3, dashed lines
represent k 5 5, and dot-dashed lines represent k 5 10.
 W. Woertman et al. / Journal of Clinical Epidemiology 66 (2013) 752e758
for the AiD trial, using a parallel group design would require
six times as much patients as the stepped wedge design. In
comparison, for an ANCOVA design, the design effect is
ð1  0:692 Þ  2:951:52, so 910 patients would be required.
This shows that the ANCOVA design also requires far less
patients than the parallel group design, but still over three
times as many patients as the stepped wedge design.
5. Discussion
We showed that, for CRTs, the stepped wedge design is
far more efficient than a parallel group design in terms of
sample size. Besides, we also showed that the stepped
wedge design is more efficient than an ANCOVA design.
This can be explained by the fact that a stepped wedge
design involves repeated measurements by definition. Fur-
thermore, all clusters receive both the control and the treat-
ment condition, so they can act as their own control.
Therefore, the treatment effect can be estimated from
between- and within-cluster comparisons.
It should be noted that the comparison is made with
a parallel group design that uses only one measurement
and an ANCOVA design that uses two measurements.
Hence, it would be more fair to make a comparison with
a repeated measurements parallel group or ANCOVA
design with an equal number of measurements as a corre-
sponding stepped wedge design. However, design effects
for such designs (with equal assumptions) are not yet avail-
able. It can be expected that a stepped wedge design will
still be more efficient than a repeated measurements paral-
lel group design because the latter design only uses
between-cluster comparisons, whereas the stepped wedge
design uses both between- and within-cluster comparisons
to estimate the treatment effect. For a repeated measure-
ments ANCOVA design, the comparison is less straightfor-
ward and therefore subject to further research.
In the introduction, we mentioned the crossover design,
so we could have questioned how the efficiency of a stepped
wedge design compares to such a design. The crossover
design is useful when carryover effects are absent. Then,
the crossover design is expected to be more efficient than
the simplest stepped wedge design with two steps because
within the crossover design two between- and two within-
cluster comparisons can be made by using two measure-
ments, whereas within the stepped wedge design only one
between- and two within-cluster comparisons can be made
by using three measurements. Yet, for larger numbers of
steps within the stepped wedge design, this design may be-
come more efficient in terms of sample size. Furthermore,
if carryover effects are present, the crossover design gives
biased results and will be less efficient than the stepped
wedge design. In our example of the AiD study, carryover
effects were very likely to be present because the program
involved training of professionals. Hence, the stepped
wedge design was preferred over a crossover design. More
757
information about the design of crossover CRTs can be
found elsewhere [6,17].
Although the stepped wedge design is more efficient in
terms of sample size, the properties of this design have
some less attractive consequences. First, the inherent re-
peated measurements lead to a longer study period, which
will come with additional costs. Besides, the costs for im-
plementation will be higher because all clusters within this
design will receive the intervention during the study even-
tually. In contrast, in a parallel group design, only half of
the clusters will receive the intervention. Furthermore, the
analysis of the data is more complex for stepped wedge
designs than for parallel group or ANCOVA designs [8,9]
because data from a stepped wedge design are correlated
because of the clustering of subjects within clusters as well
as the repeated measurements. Hence, analysis methods for
correlated data can use, for example, generalized linear
mixed models or generalized estimating equations [8].
Another implication of implementing the intervention in
all clusters sequentially is that, depending on the type of
intervention, learning effects may occur in the trainers or
appliers of the intervention. The AiD study, for example,
involved training of nurses, physicians, psychologists, and
recreational therapists to detect and treat depression. The
trainer(s) of this program may become more experienced
after training every new enrolling NH unit, which may lead
to (small) differences in the program between units. Subse-
quently, this may have an effect on the estimated treatment
effect across clusters. Besides, clusters will differ in the
amount of time spent in the treatment condition. Eventu-
ally, clusters that switch at the first step will be more expe-
rienced with the treatment than clusters that switch at later
steps. This may also affect the estimated treatment effect.
However, both types of learning effects can be modeled.
An often-mentioned implication of repeated measure-
ments is a higher burden on the respondents. In our sample
size formula, we assumed that there would be no within-
subject correlation over time. This is most likely to be true
in cross-sectional studies, when new subjects are sampled
from the clusters at each measurement. Therefore, the bur-
den of repeated measurements on the subjects will be absent.
However, not only the subjects but also the clusters are in-
volved. So, there might be burden at the cluster level. In
the AiD study, for example, a part of the outcome measures
had to be carried out by the unit personnel at each step.
Hence, a limited number of measurements is recommended.
In principle, the stepped wedge design is not only exclu-
sively applicable to CRTs but it can also be used in individ-
ually randomized trials. In fact, the individually randomized
trial equals a CRT with only one subject per cluster. The
same formulae apply, but here the burden of repeatedly
measuring the subject may become a problem. If this is
the case, it is recommended to restrict the number of steps.
The cross-sectional nature of our approach has an effect
on the statistical power. In general, cohort studies are more
powerful than cross-sectional studies [18e20]. Subjects are
758 W. Woertman et al. / Journal of Clinical Epidemiology 66 (2013) 752e758
measured repeatedly within cohort designs. Therefore, not
only within-cluster but also within-subject comparisons
can be used to estimate the treatment effect. Hence, the
required sample size for a cohort study will be smaller than
for a cross-sectional study. A sample size formula that
allows for within-subject correlations over time is not avail-
able yet. However, using our formula will at least provide
a sample size with sufficient power in case of a cohort
design.
We stated that when the number of clusters that should
switch at each step is not an integer, it suffices to distribute
the clusters as evenly as possible over the steps. Yet, this is
an approximation, that is, it is not known how the power is
affected by the uneven distribution and what if the large
groups of clusters should switch at early steps or at later
steps. For example, if seven clusters are required for three
steps, then there are two groups of two clusters and one
group of three clusters for the three steps. Then, it can be
questioned if the group of three clusters should switch at
the first, second, or third step. Practical considerations may
influence this choice. However, the effect of the uneven dis-
tribution on statistical power is a topic for further research.
In conclusion, the usefulness of the stepped wedge design
is increasingly being recognized by researchers [10]. How-
ever, a simple sample size formula for this design was lack-
ing. We presented a formula in which the number of steps,
measurements, and cluster sizes can be varied. Besides, we
showed how these choices affect the required sample size.
Hence, designing a stepped wedge CRT is simplified.
Appendix
Supplementary data
Supplementary data related to this article can be found at
http://dx.doi.org/10.1016/j.jclinepi.2013.01.009.
References
[1] Schulz KF, Altman DG, Moher D. CONSORT 2010 statement:
updated guidelines for reporting parallel group randomised trials.
J Clin Epidemiol 2010;63:834e40.
[2] Bland JM. Cluster randomised trials in the medical literature: two
bibliometric surveys. BMC Med Res Methodol 2004;4:21.
[3] Donner A, Klar N. Design and analysis of cluster randomization trials
in health research. London, UK: Arnold; 2000.
[4] ICH E9 Expert Working Group. Statistical principles for clinical
trials. Stat Med 1999;18:1905e42.
[5] Teerenstra S, Eldridge S, Graff M, de Hoop E, Borm GF. A simple
sample size formula for analysis of covariance in cluster randomized
trials. Stat Med 2012;31:2169e78.
[6] Rietbergen C, Moerbeek M. The design of cluster randomized cross-
over trials. J Educ Behav Stat 2011;36:472e90.
[7] The Gambia Hepatitis intervention study. Cancer Res
1987;47:5782e7.
[8] Hussey MA, Hughes JP. Design and analysis of stepped wedge
cluster randomized trials. Contemp Clin Trials 2007;28:182e91.
[9] Brown CA, Lilford RJ. The stepped wedge trial design: a systematic
review. BMC Med Res Methodol 2006;6:54.
[10] Mdege ND, Man MS, Taylor Nee Brown CA, Torgerson DJ. System-
atic review of stepped wedge cluster randomized trials shows that
design is particularly used to evaluate interventions during routine
implementation. J Clin Epidemiol 2011;64:936e48.
[11] Gerritsen DL, Smalbrugge M, Teerenstra S, Leontjevas R,
Adang EM, Vernooij-Dassen M, et al. Act in case of depression:
the evaluation of a care program to improve the detection and
treatment of depression in nursing homes. Study Protocol. BMC
Psychiatry 2011;11:91.
[12] Feng ZD, Diehr P, Peterson A, McLerran D. Selected statistical
issues in group randomized trials. Annu Rev Public Health
2001;22:167e87.
[13] Flynn TN, Whitley E, Peters TJ. Recruitment strategies in a cluster
randomized trialdcost implications. Stat Med 2002;21:397e405.
[14] Murray DM, Varnell SP, Blitstein JL. Design and analysis of group-
randomized trials: a review of recent methodological developments.
Am J Public Health 2004;94:423e32.
[15] Hemming K, Girling AJ, Sitch AJ, Marsh J, Lilford RJ. Sample size
calculations for cluster randomised controlled trials with a fixed
number of clusters. BMC Med Res Methodol 2011;11:102.
[16] Campbell MJ. Cluster randomized trials in general (family) practice
research. Stat Methods Med Res 2000;9:81e94.
[17] Giraudeau B, Ravaud P, Donner A. Sample size calculation for
cluster randomized cross-over trials. Stat Med 2008;27:5578e85.
[18] Duncan GJ, Kalton G. Issues of design and analysis of surveys across
time. Int Stat Rev 1987;55:97e117.
[19] Feldman HA, McKinlay SM. Cohort versus cross-sectional design in
large field trials: precision, sample size, and a unifying model. Stat
Med 1994;13:61e78.
[20] Frison L, Pocock SJ. Repeated measures in clinical-trialsdanalysis
using mean summary statistics and its implications for design. Stat
Med 1992;11:1685e704.